Forecast-driven charging optimization: Optimization for depot charging under operational uncertainty
Veikkolainen, Juho-Eemeli (2026)
2026
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-06-17
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606157483
https://urn.fi/URN:NBN:fi:tuni-202606157483
Tiivistelmä
The electrification of heavy-duty vehicle fleets has increased the need for efficient depot charging strategies that can reduce energy costs, manage peak power demand, and enable participation in electricity markets. However, practical charging optimization is challenging because detailed vehicle-level schedules and future charging requirements are often unavailable or uncertain. This thesis investigates whether economically meaningful charging optimization can be achieved without individual vehicle schedules by forecasting charging behaviour at the aggregate depot level.
The work proposes a forecast-driven optimization framework in which aggregate depot charging power is treated as an optimization variable, while future operational constraints are estimated from historical charging data. Two key forecasting targets are considered: the number of simultaneous charging sessions, which defines available charging power capacity, and aggregate future energy demand, which defines the energy that must be delivered before vehicle departures. A multi-year dataset from a commercial electric bus depot was transformed into 15-minute time series and enriched with calendar-based covariates. Several forecasting approaches were evaluated, including Prophet, a Prophet–SARIMA decomposition pipeline, XGBoost, TimesFM, and Chronos-2 in both zero-shot and fine-tuned configurations.
The results show that aggregate depot charging behaviour contains strong predictable temporal structure, but traditional additive and residual-based statistical models are insufficient for capturing its nonlinear and operationally variable patterns. Advanced machine learning and foundation models performed substantially better, with fine-tuned Chronos-2 achieving the best overall forecasting accuracy. The probabilistic forecasts further enabled quantile-based safety margins, allowing the optimization framework to balance economic performance and operational reliability. Practical evaluation indicated that the approach can capitalize on approximately 85–95% of available charging capacity while keeping around 90% of delivered energy peak- and cost-optimized. Overall, the results demonstrate that probabilistic aggregate forecasting provides a viable foundation for scalable, economically efficient, and robust depot charging optimization under operational uncertainty.
The work proposes a forecast-driven optimization framework in which aggregate depot charging power is treated as an optimization variable, while future operational constraints are estimated from historical charging data. Two key forecasting targets are considered: the number of simultaneous charging sessions, which defines available charging power capacity, and aggregate future energy demand, which defines the energy that must be delivered before vehicle departures. A multi-year dataset from a commercial electric bus depot was transformed into 15-minute time series and enriched with calendar-based covariates. Several forecasting approaches were evaluated, including Prophet, a Prophet–SARIMA decomposition pipeline, XGBoost, TimesFM, and Chronos-2 in both zero-shot and fine-tuned configurations.
The results show that aggregate depot charging behaviour contains strong predictable temporal structure, but traditional additive and residual-based statistical models are insufficient for capturing its nonlinear and operationally variable patterns. Advanced machine learning and foundation models performed substantially better, with fine-tuned Chronos-2 achieving the best overall forecasting accuracy. The probabilistic forecasts further enabled quantile-based safety margins, allowing the optimization framework to balance economic performance and operational reliability. Practical evaluation indicated that the approach can capitalize on approximately 85–95% of available charging capacity while keeping around 90% of delivered energy peak- and cost-optimized. Overall, the results demonstrate that probabilistic aggregate forecasting provides a viable foundation for scalable, economically efficient, and robust depot charging optimization under operational uncertainty.